Meta wanted real workplace behavior to train better AI agents. The problem is that its employees saw a data grab that reached deeper than mouse clicks.
Meta is already pulling back parts of its internal AI training program after weeks of staff anger over a system that records how employees use their work computers. The change matters because this is not just another workplace monitoring dispute. It is a test of how far companies will go to turn ordinary office work into training data for autonomous agents.
According to Reuters, Meta told employees on June 2 that new controls would let them pause collection for up to 30 minutes at a time and request exemptions from the initiative. The memo was written by Stephane Kasriel, a vice president in Meta's AI model-building Superintelligence Labs unit, and followed complaints about personal data, battery drain and home internet usage spikes.
The program is called the Model Capability Initiative, or MCI. Meta announced it in April as a way to collect mouse movements, clicks, keystrokes and occasional screen snapshots from U.S.-based employees on work-related apps and websites. The stated goal is straightforward: if Meta wants AI agents that can complete everyday computer tasks, those systems need examples of how people actually move through software.
That sounds reasonable in the abstract. Anyone who has watched an AI agent fail at a simple dropdown menu understands the problem. Text alone does not teach a model how a worker navigates a messy internal tool, switches between tabs, copies information into a form, corrects an error and checks whether the task is complete. Real work is full of small decisions that do not show up in a polished document.
The concern is that MCI appears to collect more than the behavior Meta initially emphasized. Internal materials reviewed by Reuters and described by TechSpot showed the system collecting interaction data across more than 200 apps and websites, with possible capture of email contents, chat messages, browsing history, clipboard actions, code changes and device activity.
Meta has said the tool is installed only on U.S. devices and is designed to analyze interaction behavior rather than the substance of communications. But even that distinction gets complicated quickly. If a U.S.-based employee has the tool enabled while emailing or chatting with a colleague outside the United States, that overseas colleague's message can still pass through the monitored device.
That is where a workplace AI project becomes a cross-border privacy issue. Meta has reportedly notified non-U.S. employees that their communications may be captured when they interact with U.S. colleagues, while telling Ireland's Data Protection Commission that collecting EU employee data is not the tool's primary purpose. The legal question is not just where the software sits. It is whether data created for work communication can be repurposed into AI training material.
For employees, the practical question is even simpler. They are being asked to help train systems that may eventually automate pieces of their own jobs. That fear did not arrive in a vacuum. Meta has been reorganizing aggressively around AI, and staff resistance has included internal criticism and posters comparing the company to an employee data extraction operation.
Why companies are tempted
Meta is not alone in chasing this kind of data. The next phase of AI competition is less about writing a clean paragraph and more about getting software to do useful work across messy interfaces. That requires training examples from real workflows, not just public web pages, synthetic tasks or neatly labeled demos.
This is why internal corporate data is becoming so valuable. Emails, chats, tickets, code reviews, documents and browser trails reveal the steps behind knowledge work. They show what people do when the policy document is outdated, when the interface changes, when the customer record has a missing field or when two internal systems disagree. That is the material AI agents need if they are going to move from assistant to operator.
But there is a difference between using employees as testers and turning their workday into a continuous training pipeline. The first can be governed with consent, limited scope, clear retention rules and visible controls. The second feels like surveillance because it captures context that workers cannot easily separate from personal judgment, private communication and job security.
Meta's new pause button and exemption process may reduce some immediate pressure, but they do not settle the larger issue. A 30-minute pause is useful if someone needs to handle sensitive material, but it also proves that employees wanted control over a system they did not believe was narrow enough in the first place.
For other technology companies, the lesson is clear. AI agents will need better behavioral training data, and internal work data is one of the richest sources available. But the companies that treat that data as a free byproduct of employment may find that the backlash arrives before the productivity gains do. What happens next at Meta will tell the market whether employee workflow data becomes a normal AI input, or a new regulatory and reputational risk that companies have to price in from the start.
Also read: AI's data center footprint is becoming a country-sized constraint • Companies are learning how to game AI search through Reddit • AI data centers are becoming a national resource test
